2 * Generic hugetlb support.
3 * (C) William Irwin, April 2004
6 #include <linux/list.h>
7 #include <linux/init.h>
8 #include <linux/module.h>
10 #include <linux/seq_file.h>
11 #include <linux/sysctl.h>
12 #include <linux/highmem.h>
13 #include <linux/mmu_notifier.h>
14 #include <linux/nodemask.h>
15 #include <linux/pagemap.h>
16 #include <linux/mempolicy.h>
17 #include <linux/cpuset.h>
18 #include <linux/mutex.h>
19 #include <linux/bootmem.h>
20 #include <linux/sysfs.h>
23 #include <asm/pgtable.h>
26 #include <linux/hugetlb.h>
29 const unsigned long hugetlb_zero = 0, hugetlb_infinity = ~0UL;
30 static gfp_t htlb_alloc_mask = GFP_HIGHUSER;
31 unsigned long hugepages_treat_as_movable;
33 static int max_hstate;
34 unsigned int default_hstate_idx;
35 struct hstate hstates[HUGE_MAX_HSTATE];
37 __initdata LIST_HEAD(huge_boot_pages);
39 /* for command line parsing */
40 static struct hstate * __initdata parsed_hstate;
41 static unsigned long __initdata default_hstate_max_huge_pages;
42 static unsigned long __initdata default_hstate_size;
44 #define for_each_hstate(h) \
45 for ((h) = hstates; (h) < &hstates[max_hstate]; (h)++)
48 * Protects updates to hugepage_freelists, nr_huge_pages, and free_huge_pages
50 static DEFINE_SPINLOCK(hugetlb_lock);
53 * Region tracking -- allows tracking of reservations and instantiated pages
54 * across the pages in a mapping.
56 * The region data structures are protected by a combination of the mmap_sem
57 * and the hugetlb_instantion_mutex. To access or modify a region the caller
58 * must either hold the mmap_sem for write, or the mmap_sem for read and
59 * the hugetlb_instantiation mutex:
61 * down_write(&mm->mmap_sem);
63 * down_read(&mm->mmap_sem);
64 * mutex_lock(&hugetlb_instantiation_mutex);
67 struct list_head link;
72 static long region_add(struct list_head *head, long f, long t)
74 struct file_region *rg, *nrg, *trg;
76 /* Locate the region we are either in or before. */
77 list_for_each_entry(rg, head, link)
81 /* Round our left edge to the current segment if it encloses us. */
85 /* Check for and consume any regions we now overlap with. */
87 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
88 if (&rg->link == head)
93 /* If this area reaches higher then extend our area to
94 * include it completely. If this is not the first area
95 * which we intend to reuse, free it. */
108 static long region_chg(struct list_head *head, long f, long t)
110 struct file_region *rg, *nrg;
113 /* Locate the region we are before or in. */
114 list_for_each_entry(rg, head, link)
118 /* If we are below the current region then a new region is required.
119 * Subtle, allocate a new region at the position but make it zero
120 * size such that we can guarantee to record the reservation. */
121 if (&rg->link == head || t < rg->from) {
122 nrg = kmalloc(sizeof(*nrg), GFP_KERNEL);
127 INIT_LIST_HEAD(&nrg->link);
128 list_add(&nrg->link, rg->link.prev);
133 /* Round our left edge to the current segment if it encloses us. */
138 /* Check for and consume any regions we now overlap with. */
139 list_for_each_entry(rg, rg->link.prev, link) {
140 if (&rg->link == head)
145 /* We overlap with this area, if it extends futher than
146 * us then we must extend ourselves. Account for its
147 * existing reservation. */
152 chg -= rg->to - rg->from;
157 static long region_truncate(struct list_head *head, long end)
159 struct file_region *rg, *trg;
162 /* Locate the region we are either in or before. */
163 list_for_each_entry(rg, head, link)
166 if (&rg->link == head)
169 /* If we are in the middle of a region then adjust it. */
170 if (end > rg->from) {
173 rg = list_entry(rg->link.next, typeof(*rg), link);
176 /* Drop any remaining regions. */
177 list_for_each_entry_safe(rg, trg, rg->link.prev, link) {
178 if (&rg->link == head)
180 chg += rg->to - rg->from;
187 static long region_count(struct list_head *head, long f, long t)
189 struct file_region *rg;
192 /* Locate each segment we overlap with, and count that overlap. */
193 list_for_each_entry(rg, head, link) {
202 seg_from = max(rg->from, f);
203 seg_to = min(rg->to, t);
205 chg += seg_to - seg_from;
212 * Convert the address within this vma to the page offset within
213 * the mapping, in pagecache page units; huge pages here.
215 static pgoff_t vma_hugecache_offset(struct hstate *h,
216 struct vm_area_struct *vma, unsigned long address)
218 return ((address - vma->vm_start) >> huge_page_shift(h)) +
219 (vma->vm_pgoff >> huge_page_order(h));
223 * Flags for MAP_PRIVATE reservations. These are stored in the bottom
224 * bits of the reservation map pointer, which are always clear due to
227 #define HPAGE_RESV_OWNER (1UL << 0)
228 #define HPAGE_RESV_UNMAPPED (1UL << 1)
229 #define HPAGE_RESV_MASK (HPAGE_RESV_OWNER | HPAGE_RESV_UNMAPPED)
232 * These helpers are used to track how many pages are reserved for
233 * faults in a MAP_PRIVATE mapping. Only the process that called mmap()
234 * is guaranteed to have their future faults succeed.
236 * With the exception of reset_vma_resv_huge_pages() which is called at fork(),
237 * the reserve counters are updated with the hugetlb_lock held. It is safe
238 * to reset the VMA at fork() time as it is not in use yet and there is no
239 * chance of the global counters getting corrupted as a result of the values.
241 * The private mapping reservation is represented in a subtly different
242 * manner to a shared mapping. A shared mapping has a region map associated
243 * with the underlying file, this region map represents the backing file
244 * pages which have ever had a reservation assigned which this persists even
245 * after the page is instantiated. A private mapping has a region map
246 * associated with the original mmap which is attached to all VMAs which
247 * reference it, this region map represents those offsets which have consumed
248 * reservation ie. where pages have been instantiated.
250 static unsigned long get_vma_private_data(struct vm_area_struct *vma)
252 return (unsigned long)vma->vm_private_data;
255 static void set_vma_private_data(struct vm_area_struct *vma,
258 vma->vm_private_data = (void *)value;
263 struct list_head regions;
266 static struct resv_map *resv_map_alloc(void)
268 struct resv_map *resv_map = kmalloc(sizeof(*resv_map), GFP_KERNEL);
272 kref_init(&resv_map->refs);
273 INIT_LIST_HEAD(&resv_map->regions);
278 static void resv_map_release(struct kref *ref)
280 struct resv_map *resv_map = container_of(ref, struct resv_map, refs);
282 /* Clear out any active regions before we release the map. */
283 region_truncate(&resv_map->regions, 0);
287 static struct resv_map *vma_resv_map(struct vm_area_struct *vma)
289 VM_BUG_ON(!is_vm_hugetlb_page(vma));
290 if (!(vma->vm_flags & VM_SHARED))
291 return (struct resv_map *)(get_vma_private_data(vma) &
296 static void set_vma_resv_map(struct vm_area_struct *vma, struct resv_map *map)
298 VM_BUG_ON(!is_vm_hugetlb_page(vma));
299 VM_BUG_ON(vma->vm_flags & VM_SHARED);
301 set_vma_private_data(vma, (get_vma_private_data(vma) &
302 HPAGE_RESV_MASK) | (unsigned long)map);
305 static void set_vma_resv_flags(struct vm_area_struct *vma, unsigned long flags)
307 VM_BUG_ON(!is_vm_hugetlb_page(vma));
308 VM_BUG_ON(vma->vm_flags & VM_SHARED);
310 set_vma_private_data(vma, get_vma_private_data(vma) | flags);
313 static int is_vma_resv_set(struct vm_area_struct *vma, unsigned long flag)
315 VM_BUG_ON(!is_vm_hugetlb_page(vma));
317 return (get_vma_private_data(vma) & flag) != 0;
320 /* Decrement the reserved pages in the hugepage pool by one */
321 static void decrement_hugepage_resv_vma(struct hstate *h,
322 struct vm_area_struct *vma)
324 if (vma->vm_flags & VM_NORESERVE)
327 if (vma->vm_flags & VM_SHARED) {
328 /* Shared mappings always use reserves */
329 h->resv_huge_pages--;
330 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
332 * Only the process that called mmap() has reserves for
335 h->resv_huge_pages--;
339 /* Reset counters to 0 and clear all HPAGE_RESV_* flags */
340 void reset_vma_resv_huge_pages(struct vm_area_struct *vma)
342 VM_BUG_ON(!is_vm_hugetlb_page(vma));
343 if (!(vma->vm_flags & VM_SHARED))
344 vma->vm_private_data = (void *)0;
347 /* Returns true if the VMA has associated reserve pages */
348 static int vma_has_reserves(struct vm_area_struct *vma)
350 if (vma->vm_flags & VM_SHARED)
352 if (is_vma_resv_set(vma, HPAGE_RESV_OWNER))
357 static void clear_gigantic_page(struct page *page,
358 unsigned long addr, unsigned long sz)
361 struct page *p = page;
364 for (i = 0; i < sz/PAGE_SIZE; i++, p = mem_map_next(p, page, i)) {
366 clear_user_highpage(p, addr + i * PAGE_SIZE);
369 static void clear_huge_page(struct page *page,
370 unsigned long addr, unsigned long sz)
374 if (unlikely(sz > MAX_ORDER_NR_PAGES))
375 return clear_gigantic_page(page, addr, sz);
378 for (i = 0; i < sz/PAGE_SIZE; i++) {
380 clear_user_highpage(page + i, addr + i * PAGE_SIZE);
384 static void copy_gigantic_page(struct page *dst, struct page *src,
385 unsigned long addr, struct vm_area_struct *vma)
388 struct hstate *h = hstate_vma(vma);
389 struct page *dst_base = dst;
390 struct page *src_base = src;
392 for (i = 0; i < pages_per_huge_page(h); ) {
394 copy_user_highpage(dst, src, addr + i*PAGE_SIZE, vma);
397 dst = mem_map_next(dst, dst_base, i);
398 src = mem_map_next(src, src_base, i);
401 static void copy_huge_page(struct page *dst, struct page *src,
402 unsigned long addr, struct vm_area_struct *vma)
405 struct hstate *h = hstate_vma(vma);
407 if (unlikely(pages_per_huge_page(h) > MAX_ORDER_NR_PAGES))
408 return copy_gigantic_page(dst, src, addr, vma);
411 for (i = 0; i < pages_per_huge_page(h); i++) {
413 copy_user_highpage(dst + i, src + i, addr + i*PAGE_SIZE, vma);
417 static void enqueue_huge_page(struct hstate *h, struct page *page)
419 int nid = page_to_nid(page);
420 list_add(&page->lru, &h->hugepage_freelists[nid]);
421 h->free_huge_pages++;
422 h->free_huge_pages_node[nid]++;
425 static struct page *dequeue_huge_page(struct hstate *h)
428 struct page *page = NULL;
430 for (nid = 0; nid < MAX_NUMNODES; ++nid) {
431 if (!list_empty(&h->hugepage_freelists[nid])) {
432 page = list_entry(h->hugepage_freelists[nid].next,
434 list_del(&page->lru);
435 h->free_huge_pages--;
436 h->free_huge_pages_node[nid]--;
443 static struct page *dequeue_huge_page_vma(struct hstate *h,
444 struct vm_area_struct *vma,
445 unsigned long address, int avoid_reserve)
448 struct page *page = NULL;
449 struct mempolicy *mpol;
450 nodemask_t *nodemask;
451 struct zonelist *zonelist = huge_zonelist(vma, address,
452 htlb_alloc_mask, &mpol, &nodemask);
457 * A child process with MAP_PRIVATE mappings created by their parent
458 * have no page reserves. This check ensures that reservations are
459 * not "stolen". The child may still get SIGKILLed
461 if (!vma_has_reserves(vma) &&
462 h->free_huge_pages - h->resv_huge_pages == 0)
465 /* If reserves cannot be used, ensure enough pages are in the pool */
466 if (avoid_reserve && h->free_huge_pages - h->resv_huge_pages == 0)
469 for_each_zone_zonelist_nodemask(zone, z, zonelist,
470 MAX_NR_ZONES - 1, nodemask) {
471 nid = zone_to_nid(zone);
472 if (cpuset_zone_allowed_softwall(zone, htlb_alloc_mask) &&
473 !list_empty(&h->hugepage_freelists[nid])) {
474 page = list_entry(h->hugepage_freelists[nid].next,
476 list_del(&page->lru);
477 h->free_huge_pages--;
478 h->free_huge_pages_node[nid]--;
481 decrement_hugepage_resv_vma(h, vma);
490 static void update_and_free_page(struct hstate *h, struct page *page)
495 h->nr_huge_pages_node[page_to_nid(page)]--;
496 for (i = 0; i < pages_per_huge_page(h); i++) {
497 page[i].flags &= ~(1 << PG_locked | 1 << PG_error | 1 << PG_referenced |
498 1 << PG_dirty | 1 << PG_active | 1 << PG_reserved |
499 1 << PG_private | 1<< PG_writeback);
501 set_compound_page_dtor(page, NULL);
502 set_page_refcounted(page);
503 arch_release_hugepage(page);
504 __free_pages(page, huge_page_order(h));
507 struct hstate *size_to_hstate(unsigned long size)
512 if (huge_page_size(h) == size)
518 static void free_huge_page(struct page *page)
521 * Can't pass hstate in here because it is called from the
522 * compound page destructor.
524 struct hstate *h = page_hstate(page);
525 int nid = page_to_nid(page);
526 struct address_space *mapping;
528 mapping = (struct address_space *) page_private(page);
529 set_page_private(page, 0);
530 BUG_ON(page_count(page));
531 INIT_LIST_HEAD(&page->lru);
533 spin_lock(&hugetlb_lock);
534 if (h->surplus_huge_pages_node[nid] && huge_page_order(h) < MAX_ORDER) {
535 update_and_free_page(h, page);
536 h->surplus_huge_pages--;
537 h->surplus_huge_pages_node[nid]--;
539 enqueue_huge_page(h, page);
541 spin_unlock(&hugetlb_lock);
543 hugetlb_put_quota(mapping, 1);
547 * Increment or decrement surplus_huge_pages. Keep node-specific counters
548 * balanced by operating on them in a round-robin fashion.
549 * Returns 1 if an adjustment was made.
551 static int adjust_pool_surplus(struct hstate *h, int delta)
557 VM_BUG_ON(delta != -1 && delta != 1);
559 nid = next_node(nid, node_online_map);
560 if (nid == MAX_NUMNODES)
561 nid = first_node(node_online_map);
563 /* To shrink on this node, there must be a surplus page */
564 if (delta < 0 && !h->surplus_huge_pages_node[nid])
566 /* Surplus cannot exceed the total number of pages */
567 if (delta > 0 && h->surplus_huge_pages_node[nid] >=
568 h->nr_huge_pages_node[nid])
571 h->surplus_huge_pages += delta;
572 h->surplus_huge_pages_node[nid] += delta;
575 } while (nid != prev_nid);
581 static void prep_new_huge_page(struct hstate *h, struct page *page, int nid)
583 set_compound_page_dtor(page, free_huge_page);
584 spin_lock(&hugetlb_lock);
586 h->nr_huge_pages_node[nid]++;
587 spin_unlock(&hugetlb_lock);
588 put_page(page); /* free it into the hugepage allocator */
591 static struct page *alloc_fresh_huge_page_node(struct hstate *h, int nid)
595 if (h->order >= MAX_ORDER)
598 page = alloc_pages_node(nid,
599 htlb_alloc_mask|__GFP_COMP|__GFP_THISNODE|
600 __GFP_REPEAT|__GFP_NOWARN,
603 if (arch_prepare_hugepage(page)) {
604 __free_pages(page, huge_page_order(h));
607 prep_new_huge_page(h, page, nid);
614 * Use a helper variable to find the next node and then
615 * copy it back to hugetlb_next_nid afterwards:
616 * otherwise there's a window in which a racer might
617 * pass invalid nid MAX_NUMNODES to alloc_pages_node.
618 * But we don't need to use a spin_lock here: it really
619 * doesn't matter if occasionally a racer chooses the
620 * same nid as we do. Move nid forward in the mask even
621 * if we just successfully allocated a hugepage so that
622 * the next caller gets hugepages on the next node.
624 static int hstate_next_node(struct hstate *h)
627 next_nid = next_node(h->hugetlb_next_nid, node_online_map);
628 if (next_nid == MAX_NUMNODES)
629 next_nid = first_node(node_online_map);
630 h->hugetlb_next_nid = next_nid;
634 static int alloc_fresh_huge_page(struct hstate *h)
641 start_nid = h->hugetlb_next_nid;
644 page = alloc_fresh_huge_page_node(h, h->hugetlb_next_nid);
647 next_nid = hstate_next_node(h);
648 } while (!page && h->hugetlb_next_nid != start_nid);
651 count_vm_event(HTLB_BUDDY_PGALLOC);
653 count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
658 static struct page *alloc_buddy_huge_page(struct hstate *h,
659 struct vm_area_struct *vma, unsigned long address)
664 if (h->order >= MAX_ORDER)
668 * Assume we will successfully allocate the surplus page to
669 * prevent racing processes from causing the surplus to exceed
672 * This however introduces a different race, where a process B
673 * tries to grow the static hugepage pool while alloc_pages() is
674 * called by process A. B will only examine the per-node
675 * counters in determining if surplus huge pages can be
676 * converted to normal huge pages in adjust_pool_surplus(). A
677 * won't be able to increment the per-node counter, until the
678 * lock is dropped by B, but B doesn't drop hugetlb_lock until
679 * no more huge pages can be converted from surplus to normal
680 * state (and doesn't try to convert again). Thus, we have a
681 * case where a surplus huge page exists, the pool is grown, and
682 * the surplus huge page still exists after, even though it
683 * should just have been converted to a normal huge page. This
684 * does not leak memory, though, as the hugepage will be freed
685 * once it is out of use. It also does not allow the counters to
686 * go out of whack in adjust_pool_surplus() as we don't modify
687 * the node values until we've gotten the hugepage and only the
688 * per-node value is checked there.
690 spin_lock(&hugetlb_lock);
691 if (h->surplus_huge_pages >= h->nr_overcommit_huge_pages) {
692 spin_unlock(&hugetlb_lock);
696 h->surplus_huge_pages++;
698 spin_unlock(&hugetlb_lock);
700 page = alloc_pages(htlb_alloc_mask|__GFP_COMP|
701 __GFP_REPEAT|__GFP_NOWARN,
704 if (page && arch_prepare_hugepage(page)) {
705 __free_pages(page, huge_page_order(h));
709 spin_lock(&hugetlb_lock);
712 * This page is now managed by the hugetlb allocator and has
713 * no users -- drop the buddy allocator's reference.
715 put_page_testzero(page);
716 VM_BUG_ON(page_count(page));
717 nid = page_to_nid(page);
718 set_compound_page_dtor(page, free_huge_page);
720 * We incremented the global counters already
722 h->nr_huge_pages_node[nid]++;
723 h->surplus_huge_pages_node[nid]++;
724 __count_vm_event(HTLB_BUDDY_PGALLOC);
727 h->surplus_huge_pages--;
728 __count_vm_event(HTLB_BUDDY_PGALLOC_FAIL);
730 spin_unlock(&hugetlb_lock);
736 * Increase the hugetlb pool such that it can accomodate a reservation
739 static int gather_surplus_pages(struct hstate *h, int delta)
741 struct list_head surplus_list;
742 struct page *page, *tmp;
744 int needed, allocated;
746 needed = (h->resv_huge_pages + delta) - h->free_huge_pages;
748 h->resv_huge_pages += delta;
753 INIT_LIST_HEAD(&surplus_list);
757 spin_unlock(&hugetlb_lock);
758 for (i = 0; i < needed; i++) {
759 page = alloc_buddy_huge_page(h, NULL, 0);
762 * We were not able to allocate enough pages to
763 * satisfy the entire reservation so we free what
764 * we've allocated so far.
766 spin_lock(&hugetlb_lock);
771 list_add(&page->lru, &surplus_list);
776 * After retaking hugetlb_lock, we need to recalculate 'needed'
777 * because either resv_huge_pages or free_huge_pages may have changed.
779 spin_lock(&hugetlb_lock);
780 needed = (h->resv_huge_pages + delta) -
781 (h->free_huge_pages + allocated);
786 * The surplus_list now contains _at_least_ the number of extra pages
787 * needed to accomodate the reservation. Add the appropriate number
788 * of pages to the hugetlb pool and free the extras back to the buddy
789 * allocator. Commit the entire reservation here to prevent another
790 * process from stealing the pages as they are added to the pool but
791 * before they are reserved.
794 h->resv_huge_pages += delta;
797 /* Free the needed pages to the hugetlb pool */
798 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
801 list_del(&page->lru);
802 enqueue_huge_page(h, page);
805 /* Free unnecessary surplus pages to the buddy allocator */
806 if (!list_empty(&surplus_list)) {
807 spin_unlock(&hugetlb_lock);
808 list_for_each_entry_safe(page, tmp, &surplus_list, lru) {
809 list_del(&page->lru);
811 * The page has a reference count of zero already, so
812 * call free_huge_page directly instead of using
813 * put_page. This must be done with hugetlb_lock
814 * unlocked which is safe because free_huge_page takes
815 * hugetlb_lock before deciding how to free the page.
817 free_huge_page(page);
819 spin_lock(&hugetlb_lock);
826 * When releasing a hugetlb pool reservation, any surplus pages that were
827 * allocated to satisfy the reservation must be explicitly freed if they were
830 static void return_unused_surplus_pages(struct hstate *h,
831 unsigned long unused_resv_pages)
835 unsigned long nr_pages;
838 * We want to release as many surplus pages as possible, spread
839 * evenly across all nodes. Iterate across all nodes until we
840 * can no longer free unreserved surplus pages. This occurs when
841 * the nodes with surplus pages have no free pages.
843 unsigned long remaining_iterations = num_online_nodes();
845 /* Uncommit the reservation */
846 h->resv_huge_pages -= unused_resv_pages;
848 /* Cannot return gigantic pages currently */
849 if (h->order >= MAX_ORDER)
852 nr_pages = min(unused_resv_pages, h->surplus_huge_pages);
854 while (remaining_iterations-- && nr_pages) {
855 nid = next_node(nid, node_online_map);
856 if (nid == MAX_NUMNODES)
857 nid = first_node(node_online_map);
859 if (!h->surplus_huge_pages_node[nid])
862 if (!list_empty(&h->hugepage_freelists[nid])) {
863 page = list_entry(h->hugepage_freelists[nid].next,
865 list_del(&page->lru);
866 update_and_free_page(h, page);
867 h->free_huge_pages--;
868 h->free_huge_pages_node[nid]--;
869 h->surplus_huge_pages--;
870 h->surplus_huge_pages_node[nid]--;
872 remaining_iterations = num_online_nodes();
878 * Determine if the huge page at addr within the vma has an associated
879 * reservation. Where it does not we will need to logically increase
880 * reservation and actually increase quota before an allocation can occur.
881 * Where any new reservation would be required the reservation change is
882 * prepared, but not committed. Once the page has been quota'd allocated
883 * an instantiated the change should be committed via vma_commit_reservation.
884 * No action is required on failure.
886 static int vma_needs_reservation(struct hstate *h,
887 struct vm_area_struct *vma, unsigned long addr)
889 struct address_space *mapping = vma->vm_file->f_mapping;
890 struct inode *inode = mapping->host;
892 if (vma->vm_flags & VM_SHARED) {
893 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
894 return region_chg(&inode->i_mapping->private_list,
897 } else if (!is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
902 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
903 struct resv_map *reservations = vma_resv_map(vma);
905 err = region_chg(&reservations->regions, idx, idx + 1);
911 static void vma_commit_reservation(struct hstate *h,
912 struct vm_area_struct *vma, unsigned long addr)
914 struct address_space *mapping = vma->vm_file->f_mapping;
915 struct inode *inode = mapping->host;
917 if (vma->vm_flags & VM_SHARED) {
918 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
919 region_add(&inode->i_mapping->private_list, idx, idx + 1);
921 } else if (is_vma_resv_set(vma, HPAGE_RESV_OWNER)) {
922 pgoff_t idx = vma_hugecache_offset(h, vma, addr);
923 struct resv_map *reservations = vma_resv_map(vma);
925 /* Mark this page used in the map. */
926 region_add(&reservations->regions, idx, idx + 1);
930 static struct page *alloc_huge_page(struct vm_area_struct *vma,
931 unsigned long addr, int avoid_reserve)
933 struct hstate *h = hstate_vma(vma);
935 struct address_space *mapping = vma->vm_file->f_mapping;
936 struct inode *inode = mapping->host;
940 * Processes that did not create the mapping will have no reserves and
941 * will not have accounted against quota. Check that the quota can be
942 * made before satisfying the allocation
943 * MAP_NORESERVE mappings may also need pages and quota allocated
944 * if no reserve mapping overlaps.
946 chg = vma_needs_reservation(h, vma, addr);
950 if (hugetlb_get_quota(inode->i_mapping, chg))
951 return ERR_PTR(-ENOSPC);
953 spin_lock(&hugetlb_lock);
954 page = dequeue_huge_page_vma(h, vma, addr, avoid_reserve);
955 spin_unlock(&hugetlb_lock);
958 page = alloc_buddy_huge_page(h, vma, addr);
960 hugetlb_put_quota(inode->i_mapping, chg);
961 return ERR_PTR(-VM_FAULT_OOM);
965 set_page_refcounted(page);
966 set_page_private(page, (unsigned long) mapping);
968 vma_commit_reservation(h, vma, addr);
973 __attribute__((weak)) int alloc_bootmem_huge_page(struct hstate *h)
975 struct huge_bootmem_page *m;
976 int nr_nodes = nodes_weight(node_online_map);
981 addr = __alloc_bootmem_node_nopanic(
982 NODE_DATA(h->hugetlb_next_nid),
983 huge_page_size(h), huge_page_size(h), 0);
987 * Use the beginning of the huge page to store the
988 * huge_bootmem_page struct (until gather_bootmem
989 * puts them into the mem_map).
1001 BUG_ON((unsigned long)virt_to_phys(m) & (huge_page_size(h) - 1));
1002 /* Put them into a private list first because mem_map is not up yet */
1003 list_add(&m->list, &huge_boot_pages);
1008 /* Put bootmem huge pages into the standard lists after mem_map is up */
1009 static void __init gather_bootmem_prealloc(void)
1011 struct huge_bootmem_page *m;
1013 list_for_each_entry(m, &huge_boot_pages, list) {
1014 struct page *page = virt_to_page(m);
1015 struct hstate *h = m->hstate;
1016 __ClearPageReserved(page);
1017 WARN_ON(page_count(page) != 1);
1018 prep_compound_page(page, h->order);
1019 prep_new_huge_page(h, page, page_to_nid(page));
1023 static void __init hugetlb_hstate_alloc_pages(struct hstate *h)
1027 for (i = 0; i < h->max_huge_pages; ++i) {
1028 if (h->order >= MAX_ORDER) {
1029 if (!alloc_bootmem_huge_page(h))
1031 } else if (!alloc_fresh_huge_page(h))
1034 h->max_huge_pages = i;
1037 static void __init hugetlb_init_hstates(void)
1041 for_each_hstate(h) {
1042 /* oversize hugepages were init'ed in early boot */
1043 if (h->order < MAX_ORDER)
1044 hugetlb_hstate_alloc_pages(h);
1048 static char * __init memfmt(char *buf, unsigned long n)
1050 if (n >= (1UL << 30))
1051 sprintf(buf, "%lu GB", n >> 30);
1052 else if (n >= (1UL << 20))
1053 sprintf(buf, "%lu MB", n >> 20);
1055 sprintf(buf, "%lu KB", n >> 10);
1059 static void __init report_hugepages(void)
1063 for_each_hstate(h) {
1065 printk(KERN_INFO "HugeTLB registered %s page size, "
1066 "pre-allocated %ld pages\n",
1067 memfmt(buf, huge_page_size(h)),
1068 h->free_huge_pages);
1072 #ifdef CONFIG_HIGHMEM
1073 static void try_to_free_low(struct hstate *h, unsigned long count)
1077 if (h->order >= MAX_ORDER)
1080 for (i = 0; i < MAX_NUMNODES; ++i) {
1081 struct page *page, *next;
1082 struct list_head *freel = &h->hugepage_freelists[i];
1083 list_for_each_entry_safe(page, next, freel, lru) {
1084 if (count >= h->nr_huge_pages)
1086 if (PageHighMem(page))
1088 list_del(&page->lru);
1089 update_and_free_page(h, page);
1090 h->free_huge_pages--;
1091 h->free_huge_pages_node[page_to_nid(page)]--;
1096 static inline void try_to_free_low(struct hstate *h, unsigned long count)
1101 #define persistent_huge_pages(h) (h->nr_huge_pages - h->surplus_huge_pages)
1102 static unsigned long set_max_huge_pages(struct hstate *h, unsigned long count)
1104 unsigned long min_count, ret;
1106 if (h->order >= MAX_ORDER)
1107 return h->max_huge_pages;
1110 * Increase the pool size
1111 * First take pages out of surplus state. Then make up the
1112 * remaining difference by allocating fresh huge pages.
1114 * We might race with alloc_buddy_huge_page() here and be unable
1115 * to convert a surplus huge page to a normal huge page. That is
1116 * not critical, though, it just means the overall size of the
1117 * pool might be one hugepage larger than it needs to be, but
1118 * within all the constraints specified by the sysctls.
1120 spin_lock(&hugetlb_lock);
1121 while (h->surplus_huge_pages && count > persistent_huge_pages(h)) {
1122 if (!adjust_pool_surplus(h, -1))
1126 while (count > persistent_huge_pages(h)) {
1128 * If this allocation races such that we no longer need the
1129 * page, free_huge_page will handle it by freeing the page
1130 * and reducing the surplus.
1132 spin_unlock(&hugetlb_lock);
1133 ret = alloc_fresh_huge_page(h);
1134 spin_lock(&hugetlb_lock);
1141 * Decrease the pool size
1142 * First return free pages to the buddy allocator (being careful
1143 * to keep enough around to satisfy reservations). Then place
1144 * pages into surplus state as needed so the pool will shrink
1145 * to the desired size as pages become free.
1147 * By placing pages into the surplus state independent of the
1148 * overcommit value, we are allowing the surplus pool size to
1149 * exceed overcommit. There are few sane options here. Since
1150 * alloc_buddy_huge_page() is checking the global counter,
1151 * though, we'll note that we're not allowed to exceed surplus
1152 * and won't grow the pool anywhere else. Not until one of the
1153 * sysctls are changed, or the surplus pages go out of use.
1155 min_count = h->resv_huge_pages + h->nr_huge_pages - h->free_huge_pages;
1156 min_count = max(count, min_count);
1157 try_to_free_low(h, min_count);
1158 while (min_count < persistent_huge_pages(h)) {
1159 struct page *page = dequeue_huge_page(h);
1162 update_and_free_page(h, page);
1164 while (count < persistent_huge_pages(h)) {
1165 if (!adjust_pool_surplus(h, 1))
1169 ret = persistent_huge_pages(h);
1170 spin_unlock(&hugetlb_lock);
1174 #define HSTATE_ATTR_RO(_name) \
1175 static struct kobj_attribute _name##_attr = __ATTR_RO(_name)
1177 #define HSTATE_ATTR(_name) \
1178 static struct kobj_attribute _name##_attr = \
1179 __ATTR(_name, 0644, _name##_show, _name##_store)
1181 static struct kobject *hugepages_kobj;
1182 static struct kobject *hstate_kobjs[HUGE_MAX_HSTATE];
1184 static struct hstate *kobj_to_hstate(struct kobject *kobj)
1187 for (i = 0; i < HUGE_MAX_HSTATE; i++)
1188 if (hstate_kobjs[i] == kobj)
1194 static ssize_t nr_hugepages_show(struct kobject *kobj,
1195 struct kobj_attribute *attr, char *buf)
1197 struct hstate *h = kobj_to_hstate(kobj);
1198 return sprintf(buf, "%lu\n", h->nr_huge_pages);
1200 static ssize_t nr_hugepages_store(struct kobject *kobj,
1201 struct kobj_attribute *attr, const char *buf, size_t count)
1204 unsigned long input;
1205 struct hstate *h = kobj_to_hstate(kobj);
1207 err = strict_strtoul(buf, 10, &input);
1211 h->max_huge_pages = set_max_huge_pages(h, input);
1215 HSTATE_ATTR(nr_hugepages);
1217 static ssize_t nr_overcommit_hugepages_show(struct kobject *kobj,
1218 struct kobj_attribute *attr, char *buf)
1220 struct hstate *h = kobj_to_hstate(kobj);
1221 return sprintf(buf, "%lu\n", h->nr_overcommit_huge_pages);
1223 static ssize_t nr_overcommit_hugepages_store(struct kobject *kobj,
1224 struct kobj_attribute *attr, const char *buf, size_t count)
1227 unsigned long input;
1228 struct hstate *h = kobj_to_hstate(kobj);
1230 err = strict_strtoul(buf, 10, &input);
1234 spin_lock(&hugetlb_lock);
1235 h->nr_overcommit_huge_pages = input;
1236 spin_unlock(&hugetlb_lock);
1240 HSTATE_ATTR(nr_overcommit_hugepages);
1242 static ssize_t free_hugepages_show(struct kobject *kobj,
1243 struct kobj_attribute *attr, char *buf)
1245 struct hstate *h = kobj_to_hstate(kobj);
1246 return sprintf(buf, "%lu\n", h->free_huge_pages);
1248 HSTATE_ATTR_RO(free_hugepages);
1250 static ssize_t resv_hugepages_show(struct kobject *kobj,
1251 struct kobj_attribute *attr, char *buf)
1253 struct hstate *h = kobj_to_hstate(kobj);
1254 return sprintf(buf, "%lu\n", h->resv_huge_pages);
1256 HSTATE_ATTR_RO(resv_hugepages);
1258 static ssize_t surplus_hugepages_show(struct kobject *kobj,
1259 struct kobj_attribute *attr, char *buf)
1261 struct hstate *h = kobj_to_hstate(kobj);
1262 return sprintf(buf, "%lu\n", h->surplus_huge_pages);
1264 HSTATE_ATTR_RO(surplus_hugepages);
1266 static struct attribute *hstate_attrs[] = {
1267 &nr_hugepages_attr.attr,
1268 &nr_overcommit_hugepages_attr.attr,
1269 &free_hugepages_attr.attr,
1270 &resv_hugepages_attr.attr,
1271 &surplus_hugepages_attr.attr,
1275 static struct attribute_group hstate_attr_group = {
1276 .attrs = hstate_attrs,
1279 static int __init hugetlb_sysfs_add_hstate(struct hstate *h)
1283 hstate_kobjs[h - hstates] = kobject_create_and_add(h->name,
1285 if (!hstate_kobjs[h - hstates])
1288 retval = sysfs_create_group(hstate_kobjs[h - hstates],
1289 &hstate_attr_group);
1291 kobject_put(hstate_kobjs[h - hstates]);
1296 static void __init hugetlb_sysfs_init(void)
1301 hugepages_kobj = kobject_create_and_add("hugepages", mm_kobj);
1302 if (!hugepages_kobj)
1305 for_each_hstate(h) {
1306 err = hugetlb_sysfs_add_hstate(h);
1308 printk(KERN_ERR "Hugetlb: Unable to add hstate %s",
1313 static void __exit hugetlb_exit(void)
1317 for_each_hstate(h) {
1318 kobject_put(hstate_kobjs[h - hstates]);
1321 kobject_put(hugepages_kobj);
1323 module_exit(hugetlb_exit);
1325 static int __init hugetlb_init(void)
1327 /* Some platform decide whether they support huge pages at boot
1328 * time. On these, such as powerpc, HPAGE_SHIFT is set to 0 when
1329 * there is no such support
1331 if (HPAGE_SHIFT == 0)
1334 if (!size_to_hstate(default_hstate_size)) {
1335 default_hstate_size = HPAGE_SIZE;
1336 if (!size_to_hstate(default_hstate_size))
1337 hugetlb_add_hstate(HUGETLB_PAGE_ORDER);
1339 default_hstate_idx = size_to_hstate(default_hstate_size) - hstates;
1340 if (default_hstate_max_huge_pages)
1341 default_hstate.max_huge_pages = default_hstate_max_huge_pages;
1343 hugetlb_init_hstates();
1345 gather_bootmem_prealloc();
1349 hugetlb_sysfs_init();
1353 module_init(hugetlb_init);
1355 /* Should be called on processing a hugepagesz=... option */
1356 void __init hugetlb_add_hstate(unsigned order)
1361 if (size_to_hstate(PAGE_SIZE << order)) {
1362 printk(KERN_WARNING "hugepagesz= specified twice, ignoring\n");
1365 BUG_ON(max_hstate >= HUGE_MAX_HSTATE);
1367 h = &hstates[max_hstate++];
1369 h->mask = ~((1ULL << (order + PAGE_SHIFT)) - 1);
1370 h->nr_huge_pages = 0;
1371 h->free_huge_pages = 0;
1372 for (i = 0; i < MAX_NUMNODES; ++i)
1373 INIT_LIST_HEAD(&h->hugepage_freelists[i]);
1374 h->hugetlb_next_nid = first_node(node_online_map);
1375 snprintf(h->name, HSTATE_NAME_LEN, "hugepages-%lukB",
1376 huge_page_size(h)/1024);
1381 static int __init hugetlb_nrpages_setup(char *s)
1384 static unsigned long *last_mhp;
1387 * !max_hstate means we haven't parsed a hugepagesz= parameter yet,
1388 * so this hugepages= parameter goes to the "default hstate".
1391 mhp = &default_hstate_max_huge_pages;
1393 mhp = &parsed_hstate->max_huge_pages;
1395 if (mhp == last_mhp) {
1396 printk(KERN_WARNING "hugepages= specified twice without "
1397 "interleaving hugepagesz=, ignoring\n");
1401 if (sscanf(s, "%lu", mhp) <= 0)
1405 * Global state is always initialized later in hugetlb_init.
1406 * But we need to allocate >= MAX_ORDER hstates here early to still
1407 * use the bootmem allocator.
1409 if (max_hstate && parsed_hstate->order >= MAX_ORDER)
1410 hugetlb_hstate_alloc_pages(parsed_hstate);
1416 __setup("hugepages=", hugetlb_nrpages_setup);
1418 static int __init hugetlb_default_setup(char *s)
1420 default_hstate_size = memparse(s, &s);
1423 __setup("default_hugepagesz=", hugetlb_default_setup);
1425 static unsigned int cpuset_mems_nr(unsigned int *array)
1428 unsigned int nr = 0;
1430 for_each_node_mask(node, cpuset_current_mems_allowed)
1436 #ifdef CONFIG_SYSCTL
1437 int hugetlb_sysctl_handler(struct ctl_table *table, int write,
1438 struct file *file, void __user *buffer,
1439 size_t *length, loff_t *ppos)
1441 struct hstate *h = &default_hstate;
1445 tmp = h->max_huge_pages;
1448 table->maxlen = sizeof(unsigned long);
1449 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1452 h->max_huge_pages = set_max_huge_pages(h, tmp);
1457 int hugetlb_treat_movable_handler(struct ctl_table *table, int write,
1458 struct file *file, void __user *buffer,
1459 size_t *length, loff_t *ppos)
1461 proc_dointvec(table, write, file, buffer, length, ppos);
1462 if (hugepages_treat_as_movable)
1463 htlb_alloc_mask = GFP_HIGHUSER_MOVABLE;
1465 htlb_alloc_mask = GFP_HIGHUSER;
1469 int hugetlb_overcommit_handler(struct ctl_table *table, int write,
1470 struct file *file, void __user *buffer,
1471 size_t *length, loff_t *ppos)
1473 struct hstate *h = &default_hstate;
1477 tmp = h->nr_overcommit_huge_pages;
1480 table->maxlen = sizeof(unsigned long);
1481 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
1484 spin_lock(&hugetlb_lock);
1485 h->nr_overcommit_huge_pages = tmp;
1486 spin_unlock(&hugetlb_lock);
1492 #endif /* CONFIG_SYSCTL */
1494 void hugetlb_report_meminfo(struct seq_file *m)
1496 struct hstate *h = &default_hstate;
1498 "HugePages_Total: %5lu\n"
1499 "HugePages_Free: %5lu\n"
1500 "HugePages_Rsvd: %5lu\n"
1501 "HugePages_Surp: %5lu\n"
1502 "Hugepagesize: %8lu kB\n",
1506 h->surplus_huge_pages,
1507 1UL << (huge_page_order(h) + PAGE_SHIFT - 10));
1510 int hugetlb_report_node_meminfo(int nid, char *buf)
1512 struct hstate *h = &default_hstate;
1514 "Node %d HugePages_Total: %5u\n"
1515 "Node %d HugePages_Free: %5u\n"
1516 "Node %d HugePages_Surp: %5u\n",
1517 nid, h->nr_huge_pages_node[nid],
1518 nid, h->free_huge_pages_node[nid],
1519 nid, h->surplus_huge_pages_node[nid]);
1522 /* Return the number pages of memory we physically have, in PAGE_SIZE units. */
1523 unsigned long hugetlb_total_pages(void)
1525 struct hstate *h = &default_hstate;
1526 return h->nr_huge_pages * pages_per_huge_page(h);
1529 static int hugetlb_acct_memory(struct hstate *h, long delta)
1533 spin_lock(&hugetlb_lock);
1535 * When cpuset is configured, it breaks the strict hugetlb page
1536 * reservation as the accounting is done on a global variable. Such
1537 * reservation is completely rubbish in the presence of cpuset because
1538 * the reservation is not checked against page availability for the
1539 * current cpuset. Application can still potentially OOM'ed by kernel
1540 * with lack of free htlb page in cpuset that the task is in.
1541 * Attempt to enforce strict accounting with cpuset is almost
1542 * impossible (or too ugly) because cpuset is too fluid that
1543 * task or memory node can be dynamically moved between cpusets.
1545 * The change of semantics for shared hugetlb mapping with cpuset is
1546 * undesirable. However, in order to preserve some of the semantics,
1547 * we fall back to check against current free page availability as
1548 * a best attempt and hopefully to minimize the impact of changing
1549 * semantics that cpuset has.
1552 if (gather_surplus_pages(h, delta) < 0)
1555 if (delta > cpuset_mems_nr(h->free_huge_pages_node)) {
1556 return_unused_surplus_pages(h, delta);
1563 return_unused_surplus_pages(h, (unsigned long) -delta);
1566 spin_unlock(&hugetlb_lock);
1570 static void hugetlb_vm_op_open(struct vm_area_struct *vma)
1572 struct resv_map *reservations = vma_resv_map(vma);
1575 * This new VMA should share its siblings reservation map if present.
1576 * The VMA will only ever have a valid reservation map pointer where
1577 * it is being copied for another still existing VMA. As that VMA
1578 * has a reference to the reservation map it cannot dissappear until
1579 * after this open call completes. It is therefore safe to take a
1580 * new reference here without additional locking.
1583 kref_get(&reservations->refs);
1586 static void hugetlb_vm_op_close(struct vm_area_struct *vma)
1588 struct hstate *h = hstate_vma(vma);
1589 struct resv_map *reservations = vma_resv_map(vma);
1590 unsigned long reserve;
1591 unsigned long start;
1595 start = vma_hugecache_offset(h, vma, vma->vm_start);
1596 end = vma_hugecache_offset(h, vma, vma->vm_end);
1598 reserve = (end - start) -
1599 region_count(&reservations->regions, start, end);
1601 kref_put(&reservations->refs, resv_map_release);
1604 hugetlb_acct_memory(h, -reserve);
1605 hugetlb_put_quota(vma->vm_file->f_mapping, reserve);
1611 * We cannot handle pagefaults against hugetlb pages at all. They cause
1612 * handle_mm_fault() to try to instantiate regular-sized pages in the
1613 * hugegpage VMA. do_page_fault() is supposed to trap this, so BUG is we get
1616 static int hugetlb_vm_op_fault(struct vm_area_struct *vma, struct vm_fault *vmf)
1622 struct vm_operations_struct hugetlb_vm_ops = {
1623 .fault = hugetlb_vm_op_fault,
1624 .open = hugetlb_vm_op_open,
1625 .close = hugetlb_vm_op_close,
1628 static pte_t make_huge_pte(struct vm_area_struct *vma, struct page *page,
1635 pte_mkwrite(pte_mkdirty(mk_pte(page, vma->vm_page_prot)));
1637 entry = huge_pte_wrprotect(mk_pte(page, vma->vm_page_prot));
1639 entry = pte_mkyoung(entry);
1640 entry = pte_mkhuge(entry);
1645 static void set_huge_ptep_writable(struct vm_area_struct *vma,
1646 unsigned long address, pte_t *ptep)
1650 entry = pte_mkwrite(pte_mkdirty(huge_ptep_get(ptep)));
1651 if (huge_ptep_set_access_flags(vma, address, ptep, entry, 1)) {
1652 update_mmu_cache(vma, address, entry);
1657 int copy_hugetlb_page_range(struct mm_struct *dst, struct mm_struct *src,
1658 struct vm_area_struct *vma)
1660 pte_t *src_pte, *dst_pte, entry;
1661 struct page *ptepage;
1664 struct hstate *h = hstate_vma(vma);
1665 unsigned long sz = huge_page_size(h);
1667 cow = (vma->vm_flags & (VM_SHARED | VM_MAYWRITE)) == VM_MAYWRITE;
1669 for (addr = vma->vm_start; addr < vma->vm_end; addr += sz) {
1670 src_pte = huge_pte_offset(src, addr);
1673 dst_pte = huge_pte_alloc(dst, addr, sz);
1677 /* If the pagetables are shared don't copy or take references */
1678 if (dst_pte == src_pte)
1681 spin_lock(&dst->page_table_lock);
1682 spin_lock_nested(&src->page_table_lock, SINGLE_DEPTH_NESTING);
1683 if (!huge_pte_none(huge_ptep_get(src_pte))) {
1685 huge_ptep_set_wrprotect(src, addr, src_pte);
1686 entry = huge_ptep_get(src_pte);
1687 ptepage = pte_page(entry);
1689 set_huge_pte_at(dst, addr, dst_pte, entry);
1691 spin_unlock(&src->page_table_lock);
1692 spin_unlock(&dst->page_table_lock);
1700 void __unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1701 unsigned long end, struct page *ref_page)
1703 struct mm_struct *mm = vma->vm_mm;
1704 unsigned long address;
1709 struct hstate *h = hstate_vma(vma);
1710 unsigned long sz = huge_page_size(h);
1713 * A page gathering list, protected by per file i_mmap_lock. The
1714 * lock is used to avoid list corruption from multiple unmapping
1715 * of the same page since we are using page->lru.
1717 LIST_HEAD(page_list);
1719 WARN_ON(!is_vm_hugetlb_page(vma));
1720 BUG_ON(start & ~huge_page_mask(h));
1721 BUG_ON(end & ~huge_page_mask(h));
1723 mmu_notifier_invalidate_range_start(mm, start, end);
1724 spin_lock(&mm->page_table_lock);
1725 for (address = start; address < end; address += sz) {
1726 ptep = huge_pte_offset(mm, address);
1730 if (huge_pmd_unshare(mm, &address, ptep))
1734 * If a reference page is supplied, it is because a specific
1735 * page is being unmapped, not a range. Ensure the page we
1736 * are about to unmap is the actual page of interest.
1739 pte = huge_ptep_get(ptep);
1740 if (huge_pte_none(pte))
1742 page = pte_page(pte);
1743 if (page != ref_page)
1747 * Mark the VMA as having unmapped its page so that
1748 * future faults in this VMA will fail rather than
1749 * looking like data was lost
1751 set_vma_resv_flags(vma, HPAGE_RESV_UNMAPPED);
1754 pte = huge_ptep_get_and_clear(mm, address, ptep);
1755 if (huge_pte_none(pte))
1758 page = pte_page(pte);
1760 set_page_dirty(page);
1761 list_add(&page->lru, &page_list);
1763 spin_unlock(&mm->page_table_lock);
1764 flush_tlb_range(vma, start, end);
1765 mmu_notifier_invalidate_range_end(mm, start, end);
1766 list_for_each_entry_safe(page, tmp, &page_list, lru) {
1767 list_del(&page->lru);
1772 void unmap_hugepage_range(struct vm_area_struct *vma, unsigned long start,
1773 unsigned long end, struct page *ref_page)
1775 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
1776 __unmap_hugepage_range(vma, start, end, ref_page);
1777 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
1781 * This is called when the original mapper is failing to COW a MAP_PRIVATE
1782 * mappping it owns the reserve page for. The intention is to unmap the page
1783 * from other VMAs and let the children be SIGKILLed if they are faulting the
1786 static int unmap_ref_private(struct mm_struct *mm, struct vm_area_struct *vma,
1787 struct page *page, unsigned long address)
1789 struct vm_area_struct *iter_vma;
1790 struct address_space *mapping;
1791 struct prio_tree_iter iter;
1795 * vm_pgoff is in PAGE_SIZE units, hence the different calculation
1796 * from page cache lookup which is in HPAGE_SIZE units.
1798 address = address & huge_page_mask(hstate_vma(vma));
1799 pgoff = ((address - vma->vm_start) >> PAGE_SHIFT)
1800 + (vma->vm_pgoff >> PAGE_SHIFT);
1801 mapping = (struct address_space *)page_private(page);
1803 vma_prio_tree_foreach(iter_vma, &iter, &mapping->i_mmap, pgoff, pgoff) {
1804 /* Do not unmap the current VMA */
1805 if (iter_vma == vma)
1809 * Unmap the page from other VMAs without their own reserves.
1810 * They get marked to be SIGKILLed if they fault in these
1811 * areas. This is because a future no-page fault on this VMA
1812 * could insert a zeroed page instead of the data existing
1813 * from the time of fork. This would look like data corruption
1815 if (!is_vma_resv_set(iter_vma, HPAGE_RESV_OWNER))
1816 unmap_hugepage_range(iter_vma,
1817 address, address + HPAGE_SIZE,
1824 static int hugetlb_cow(struct mm_struct *mm, struct vm_area_struct *vma,
1825 unsigned long address, pte_t *ptep, pte_t pte,
1826 struct page *pagecache_page)
1828 struct hstate *h = hstate_vma(vma);
1829 struct page *old_page, *new_page;
1831 int outside_reserve = 0;
1833 old_page = pte_page(pte);
1836 /* If no-one else is actually using this page, avoid the copy
1837 * and just make the page writable */
1838 avoidcopy = (page_count(old_page) == 1);
1840 set_huge_ptep_writable(vma, address, ptep);
1845 * If the process that created a MAP_PRIVATE mapping is about to
1846 * perform a COW due to a shared page count, attempt to satisfy
1847 * the allocation without using the existing reserves. The pagecache
1848 * page is used to determine if the reserve at this address was
1849 * consumed or not. If reserves were used, a partial faulted mapping
1850 * at the time of fork() could consume its reserves on COW instead
1851 * of the full address range.
1853 if (!(vma->vm_flags & VM_SHARED) &&
1854 is_vma_resv_set(vma, HPAGE_RESV_OWNER) &&
1855 old_page != pagecache_page)
1856 outside_reserve = 1;
1858 page_cache_get(old_page);
1859 new_page = alloc_huge_page(vma, address, outside_reserve);
1861 if (IS_ERR(new_page)) {
1862 page_cache_release(old_page);
1865 * If a process owning a MAP_PRIVATE mapping fails to COW,
1866 * it is due to references held by a child and an insufficient
1867 * huge page pool. To guarantee the original mappers
1868 * reliability, unmap the page from child processes. The child
1869 * may get SIGKILLed if it later faults.
1871 if (outside_reserve) {
1872 BUG_ON(huge_pte_none(pte));
1873 if (unmap_ref_private(mm, vma, old_page, address)) {
1874 BUG_ON(page_count(old_page) != 1);
1875 BUG_ON(huge_pte_none(pte));
1876 goto retry_avoidcopy;
1881 return -PTR_ERR(new_page);
1884 spin_unlock(&mm->page_table_lock);
1885 copy_huge_page(new_page, old_page, address, vma);
1886 __SetPageUptodate(new_page);
1887 spin_lock(&mm->page_table_lock);
1889 ptep = huge_pte_offset(mm, address & huge_page_mask(h));
1890 if (likely(pte_same(huge_ptep_get(ptep), pte))) {
1892 huge_ptep_clear_flush(vma, address, ptep);
1893 set_huge_pte_at(mm, address, ptep,
1894 make_huge_pte(vma, new_page, 1));
1895 /* Make the old page be freed below */
1896 new_page = old_page;
1898 page_cache_release(new_page);
1899 page_cache_release(old_page);
1903 /* Return the pagecache page at a given address within a VMA */
1904 static struct page *hugetlbfs_pagecache_page(struct hstate *h,
1905 struct vm_area_struct *vma, unsigned long address)
1907 struct address_space *mapping;
1910 mapping = vma->vm_file->f_mapping;
1911 idx = vma_hugecache_offset(h, vma, address);
1913 return find_lock_page(mapping, idx);
1916 static int hugetlb_no_page(struct mm_struct *mm, struct vm_area_struct *vma,
1917 unsigned long address, pte_t *ptep, int write_access)
1919 struct hstate *h = hstate_vma(vma);
1920 int ret = VM_FAULT_SIGBUS;
1924 struct address_space *mapping;
1928 * Currently, we are forced to kill the process in the event the
1929 * original mapper has unmapped pages from the child due to a failed
1930 * COW. Warn that such a situation has occured as it may not be obvious
1932 if (is_vma_resv_set(vma, HPAGE_RESV_UNMAPPED)) {
1934 "PID %d killed due to inadequate hugepage pool\n",
1939 mapping = vma->vm_file->f_mapping;
1940 idx = vma_hugecache_offset(h, vma, address);
1943 * Use page lock to guard against racing truncation
1944 * before we get page_table_lock.
1947 page = find_lock_page(mapping, idx);
1949 size = i_size_read(mapping->host) >> huge_page_shift(h);
1952 page = alloc_huge_page(vma, address, 0);
1954 ret = -PTR_ERR(page);
1957 clear_huge_page(page, address, huge_page_size(h));
1958 __SetPageUptodate(page);
1960 if (vma->vm_flags & VM_SHARED) {
1962 struct inode *inode = mapping->host;
1964 err = add_to_page_cache(page, mapping, idx, GFP_KERNEL);
1972 spin_lock(&inode->i_lock);
1973 inode->i_blocks += blocks_per_huge_page(h);
1974 spin_unlock(&inode->i_lock);
1980 * If we are going to COW a private mapping later, we examine the
1981 * pending reservations for this page now. This will ensure that
1982 * any allocations necessary to record that reservation occur outside
1985 if (write_access && !(vma->vm_flags & VM_SHARED))
1986 if (vma_needs_reservation(h, vma, address) < 0) {
1988 goto backout_unlocked;
1991 spin_lock(&mm->page_table_lock);
1992 size = i_size_read(mapping->host) >> huge_page_shift(h);
1997 if (!huge_pte_none(huge_ptep_get(ptep)))
2000 new_pte = make_huge_pte(vma, page, ((vma->vm_flags & VM_WRITE)
2001 && (vma->vm_flags & VM_SHARED)));
2002 set_huge_pte_at(mm, address, ptep, new_pte);
2004 if (write_access && !(vma->vm_flags & VM_SHARED)) {
2005 /* Optimization, do the COW without a second fault */
2006 ret = hugetlb_cow(mm, vma, address, ptep, new_pte, page);
2009 spin_unlock(&mm->page_table_lock);
2015 spin_unlock(&mm->page_table_lock);
2022 int hugetlb_fault(struct mm_struct *mm, struct vm_area_struct *vma,
2023 unsigned long address, int write_access)
2028 struct page *pagecache_page = NULL;
2029 static DEFINE_MUTEX(hugetlb_instantiation_mutex);
2030 struct hstate *h = hstate_vma(vma);
2032 ptep = huge_pte_alloc(mm, address, huge_page_size(h));
2034 return VM_FAULT_OOM;
2037 * Serialize hugepage allocation and instantiation, so that we don't
2038 * get spurious allocation failures if two CPUs race to instantiate
2039 * the same page in the page cache.
2041 mutex_lock(&hugetlb_instantiation_mutex);
2042 entry = huge_ptep_get(ptep);
2043 if (huge_pte_none(entry)) {
2044 ret = hugetlb_no_page(mm, vma, address, ptep, write_access);
2051 * If we are going to COW the mapping later, we examine the pending
2052 * reservations for this page now. This will ensure that any
2053 * allocations necessary to record that reservation occur outside the
2054 * spinlock. For private mappings, we also lookup the pagecache
2055 * page now as it is used to determine if a reservation has been
2058 if (write_access && !pte_write(entry)) {
2059 if (vma_needs_reservation(h, vma, address) < 0) {
2064 if (!(vma->vm_flags & VM_SHARED))
2065 pagecache_page = hugetlbfs_pagecache_page(h,
2069 spin_lock(&mm->page_table_lock);
2070 /* Check for a racing update before calling hugetlb_cow */
2071 if (unlikely(!pte_same(entry, huge_ptep_get(ptep))))
2072 goto out_page_table_lock;
2076 if (!pte_write(entry)) {
2077 ret = hugetlb_cow(mm, vma, address, ptep, entry,
2079 goto out_page_table_lock;
2081 entry = pte_mkdirty(entry);
2083 entry = pte_mkyoung(entry);
2084 if (huge_ptep_set_access_flags(vma, address, ptep, entry, write_access))
2085 update_mmu_cache(vma, address, entry);
2087 out_page_table_lock:
2088 spin_unlock(&mm->page_table_lock);
2090 if (pagecache_page) {
2091 unlock_page(pagecache_page);
2092 put_page(pagecache_page);
2096 mutex_unlock(&hugetlb_instantiation_mutex);
2101 /* Can be overriden by architectures */
2102 __attribute__((weak)) struct page *
2103 follow_huge_pud(struct mm_struct *mm, unsigned long address,
2104 pud_t *pud, int write)
2110 static int huge_zeropage_ok(pte_t *ptep, int write, int shared)
2112 if (!ptep || write || shared)
2115 return huge_pte_none(huge_ptep_get(ptep));
2118 int follow_hugetlb_page(struct mm_struct *mm, struct vm_area_struct *vma,
2119 struct page **pages, struct vm_area_struct **vmas,
2120 unsigned long *position, int *length, int i,
2123 unsigned long pfn_offset;
2124 unsigned long vaddr = *position;
2125 int remainder = *length;
2126 struct hstate *h = hstate_vma(vma);
2127 int zeropage_ok = 0;
2128 int shared = vma->vm_flags & VM_SHARED;
2130 spin_lock(&mm->page_table_lock);
2131 while (vaddr < vma->vm_end && remainder) {
2136 * Some archs (sparc64, sh*) have multiple pte_ts to
2137 * each hugepage. We have to make * sure we get the
2138 * first, for the page indexing below to work.
2140 pte = huge_pte_offset(mm, vaddr & huge_page_mask(h));
2141 if (huge_zeropage_ok(pte, write, shared))
2145 (huge_pte_none(huge_ptep_get(pte)) && !zeropage_ok) ||
2146 (write && !pte_write(huge_ptep_get(pte)))) {
2149 spin_unlock(&mm->page_table_lock);
2150 ret = hugetlb_fault(mm, vma, vaddr, write);
2151 spin_lock(&mm->page_table_lock);
2152 if (!(ret & VM_FAULT_ERROR))
2161 pfn_offset = (vaddr & ~huge_page_mask(h)) >> PAGE_SHIFT;
2162 page = pte_page(huge_ptep_get(pte));
2166 pages[i] = ZERO_PAGE(0);
2168 pages[i] = mem_map_offset(page, pfn_offset);
2179 if (vaddr < vma->vm_end && remainder &&
2180 pfn_offset < pages_per_huge_page(h)) {
2182 * We use pfn_offset to avoid touching the pageframes
2183 * of this compound page.
2188 spin_unlock(&mm->page_table_lock);
2189 *length = remainder;
2195 void hugetlb_change_protection(struct vm_area_struct *vma,
2196 unsigned long address, unsigned long end, pgprot_t newprot)
2198 struct mm_struct *mm = vma->vm_mm;
2199 unsigned long start = address;
2202 struct hstate *h = hstate_vma(vma);
2204 BUG_ON(address >= end);
2205 flush_cache_range(vma, address, end);
2207 spin_lock(&vma->vm_file->f_mapping->i_mmap_lock);
2208 spin_lock(&mm->page_table_lock);
2209 for (; address < end; address += huge_page_size(h)) {
2210 ptep = huge_pte_offset(mm, address);
2213 if (huge_pmd_unshare(mm, &address, ptep))
2215 if (!huge_pte_none(huge_ptep_get(ptep))) {
2216 pte = huge_ptep_get_and_clear(mm, address, ptep);
2217 pte = pte_mkhuge(pte_modify(pte, newprot));
2218 set_huge_pte_at(mm, address, ptep, pte);
2221 spin_unlock(&mm->page_table_lock);
2222 spin_unlock(&vma->vm_file->f_mapping->i_mmap_lock);
2224 flush_tlb_range(vma, start, end);
2227 int hugetlb_reserve_pages(struct inode *inode,
2229 struct vm_area_struct *vma)
2232 struct hstate *h = hstate_inode(inode);
2234 if (vma && vma->vm_flags & VM_NORESERVE)
2238 * Shared mappings base their reservation on the number of pages that
2239 * are already allocated on behalf of the file. Private mappings need
2240 * to reserve the full area even if read-only as mprotect() may be
2241 * called to make the mapping read-write. Assume !vma is a shm mapping
2243 if (!vma || vma->vm_flags & VM_SHARED)
2244 chg = region_chg(&inode->i_mapping->private_list, from, to);
2246 struct resv_map *resv_map = resv_map_alloc();
2252 set_vma_resv_map(vma, resv_map);
2253 set_vma_resv_flags(vma, HPAGE_RESV_OWNER);
2259 if (hugetlb_get_quota(inode->i_mapping, chg))
2261 ret = hugetlb_acct_memory(h, chg);
2263 hugetlb_put_quota(inode->i_mapping, chg);
2266 if (!vma || vma->vm_flags & VM_SHARED)
2267 region_add(&inode->i_mapping->private_list, from, to);
2271 void hugetlb_unreserve_pages(struct inode *inode, long offset, long freed)
2273 struct hstate *h = hstate_inode(inode);
2274 long chg = region_truncate(&inode->i_mapping->private_list, offset);
2276 spin_lock(&inode->i_lock);
2277 inode->i_blocks -= blocks_per_huge_page(h);
2278 spin_unlock(&inode->i_lock);
2280 hugetlb_put_quota(inode->i_mapping, (chg - freed));
2281 hugetlb_acct_memory(h, -(chg - freed));